Tagged AI chip startups · show all articles
This cohort has profiled more than a dozen AI-chip startups individually. Zoomed out, they form a single mapped landscape — and the map itself tells a story none of the individual profiles can.
Nearly every AI accelerator in this cohort runs on a closed, proprietary architecture. Tenstorrent is building on RISC-V, an open standard — a bet that openness itself is a competitive advantage.
While most of this cohort tracks gigawatt-scale datacenter deals, SiMa.ai is building AI chips meant to run inside a camera, a robot, or a car — nowhere near a datacenter at all.
Most accelerators run a fixed circuit and make the software adapt to it. SambaNova's reconfigurable dataflow architecture does the opposite — reshaping the hardware itself around whatever model it's running.
Korea's chip industry is world-famous for memory. Rebellions is trying to build the country's own answer to Nvidia — a very different bet than anything SK hynix or Samsung's memory business is making.
Every accelerator covered elsewhere in this cohort, however specialized, still computes digitally. Rain AI is building analog, brain-inspired silicon instead — the single highest-risk, highest-upside bet in this cohort's entire startup track.
Most accelerator startups covered in this cohort hedge toward flexibility. Positron went the other way — a chip that does inference and nothing else, betting that specialization beats generality even against Nvidia.
Nvidia and Marvell are both betting on optics to connect chips together. Lightmatter's bet goes one step further: using light for the connections inside a chip package as well.
Groq built its whole identity on the claim that GPUs are the wrong chip for AI inference. Then Nvidia got involved. Here's the short version.
Every other accelerator in this cohort is designed to run whatever model architecture a customer throws at it. Etched's chip runs transformers and nothing else — an enormous bet that the architecture behind ChatGPT isn't going anywhere.